Breaking News
News & Analysis

TSMC, IBM Detail 7-nm Work

Separate talks upbeat on EUV lithography
12/6/2016 10:30 AM EST
12 comments
NO RATINGS
Page 1 / 2 Next >
More Related Links
View Comments: Newest First | Oldest First | Threaded View
Page 1 / 2   >   >>
witeken
User Rank
Freelancer
Re: scaling slow down
witeken   12/7/2016 1:52:02 PM
NO RATINGS
Okay. But a fin does not have a variable fin width. It's been at 8nm for both 22nm and 14nm.

Intel has shown a graph at IDF that they expect power ocnsumption at 10nm to come down much more than historically. That just shows how variable these things are, although the combined performance+power improvements are pretty consistent accross the nodes.

AKH0
User Rank
Author
Re: scaling slow down
AKH0   12/7/2016 12:56:35 PM
NO RATINGS
I am not talking about Dennard's prescription of what ion implant dose is needed or how all dimensions and operating voltage has to scale. At the heart of those equation is physics law that both current and capacitance scale with linear scaling if everything else is kept same; that is performance is CV/I and active power is I.V.  Going node to node the total transistor width scales by 30% if currennt density (A/um the way reported in IEDM) stays constant and so does the total cap assuming that metal thickness scales at the same rate that metal spacing scales. Yes, not everything scales at the same rate and that's where innovations come e.g. with Cu metal/low-k. A better transistor (e.g. FinFET, adding high-k, adding strain, etc) is expected to do a better than 30% per node.

Gate capacitance has not been the major contributor to the total capacitance for over a decade. So the fact that Lg did or did not scale does not change total capaictance. Drive current is also not a strong function of Lg once you get to velcoity saturation or source side injection (whether ballistic or non ballistic) whichever model you favor.

Quantum effects were playing a rule for a long time. even at 90nm with the physical oxide thickness of 1.2nm, you have a 0.3-0.4nm adder due to quantum effects.

witeken
User Rank
Freelancer
Re: scaling slow down
witeken   12/7/2016 12:04:47 PM
NO RATINGS
"Says who?"

Says quantum mechanics. Says the fact that transistor scaling hasn't been Dennard-like for well over a decande. Says the fact that certain important features of the transistor, like Lgate and especially the dielectric insulator, have stopped scaling in the last decade. Says the fact that a transistor isn't a Dennard-transistor anymore but a 3D finFET. Says the fact that Dennard scaling completely ignores the interconnect.

Moore's Law is an economic law and has nothing to do with Dennard scaling.

So I repeat: a modern quantum mechanical transistor is much, much too complicated to be approximated by the classical Dennard model.

AKH0
User Rank
Author
Re: scaling slow down
AKH0   12/7/2016 9:57:05 AM
NO RATINGS
Says who? This is simple geometry scaling. Traditional Dennard's scaling also required scaling of Vdd at the same pace. I see that part getting a hit because of Vt not scaling as fast. If scaling does not translate into expected power reduction (Dennard's rule) and cost per transistor reduction (Moore's law) there is absolutely no point in doing it

witeken
User Rank
Freelancer
Re: scaling slow down
witeken   12/7/2016 9:50:44 AM
NO RATINGS
Dennard's Law is dead. I don't understand why you're referring to it.

AKH0
User Rank
Author
Re: scaling slow down
AKH0   12/7/2016 9:03:59 AM
NO RATINGS
The 65% SRAM power reduction over two nodes might come along or even better than expected Dennard's rule. But SRAM bitcell is not limited by metal pitch so doesn't get a hit from metal-metal cap increase.

Lefty4
User Rank
Rookie
Re: scaling slow down
Lefty4   12/7/2016 9:03:41 AM
NO RATINGS
>Wondering where 12nm fits into this trend.

if it even exists... bare in mind that 12nm is only a rumour. TSMC has not officially announced it yet.

AKH0
User Rank
Author
Re: scaling slow down
AKH0   12/7/2016 8:58:42 AM
NO RATINGS
I would settle even for a 0.6x scaling per node if there were power saving (isoperformance) beyond the 30% per node that comes with Dennard's scaling. But this is in a mere 0.75x per node.

Violoncelles
User Rank
Author
Re: scaling slow down
Violoncelles   12/7/2016 7:58:35 AM
NO RATINGS
Usually , one node is deduced from the previous one by dividing by the square root of 2. So from 28 nm , we have 20 nm , 14 nm , 10 nm and 7 nm. 

resistion
User Rank
Author
Albany
resistion   12/7/2016 1:01:54 AM
NO RATINGS
The IBM alliance paper was done at Albany. The density was noticeably looser in one direction, probably resulting from their patterning choice(s).

Page 1 / 2   >   >>
Most Recent Comments

Datasheets.com Parts Search

185 million searchable parts
(please enter a part number or hit search to begin)
Like Us on Facebook
EE Times on Twitter
EE Times Twitter Feed